Breakup of the polar stratospheric vortex in the Northern Hemisphere is an event that is known to be predictable for up to a week or so ahead. This is illustrated using data from the 45-yr ECMWF Re-Analysis (ERA-40) for the sudden warmings of January 1958 and February 1979 and operational ECMWF data for February 2003. It is then shown that a similar level of skill was achieved in operational forecasts for the split of the southern stratospheric vortex in late September 2002. The highly unusual flow conditions nevertheless exposed a computational instability of the forecast model. Analyses and forecasts from reruns using improved versions of the forecasting system are presented. Isentropic maps of potential vorticity and specific humidity provide striking pictures of the advective processes at work. Forecasts as well as analyses are shown to be in good agreement with radiosonde measurements of the temperature changes associated with vortex movement, distortion, and breakup during August and September. Forecasts from 17 September onward capture the remarkable temperature rise of about 60°C recorded at 20 hPa by the Halley radiosonde station as the vortex split. Objective forecast verification and data denial experiments are used to characterize the performance of the observing and data assimilation systems and to infer overall forecast, analysis, and observation accuracy. The observations and analyses from 1957 onward in the ERA-40 archive confirm the extreme nature of the 2002 event. Secondary vortex development by barotropic instability is also discussed; in analyses for early October 2002, the process is active in the breakup of the weaker of the two vortices formed by the late-September split.
Toward the end of September 2002, the cold polar vortex in the Southern Hemisphere stratosphere elongated and split in a manner similar to that seen every few years in the Northern Hemisphere stratosphere but never before observed in the Southern Hemisphere. Despite its rarity, the event was predicted accurately up to about a week in advance by the operational forecasting system of the European Centre for Medium-Range Weather Forecasts (ECMWF). Although the forecasts were successful, the unusual flow conditions exposed a weakness in the numerical stability of the forecast model. Improved results, in terms of both reduced noise and slightly increased forecast accuracy, have been obtained in reruns of the data assimilation and forecasts that span the period from the beginning of August to mid-October. The resulting high-resolution reanalyses provide a resource for future studies of the behavior of the Antarctic stratosphere during the exceptional late winter and early spring of 2002. The impact of different components of the observing system on analysis and forecast accuracy has been studied in additional runs of the forecasting system for this period.
In this paper, we illustrate and discuss the performance of the ECMWF forecasting system for this event and use the analyses and observations from ECMWF’s latest long-term reanalysis project, the 45-yr ECMWF Re-Analysis (ERA-40), to help put the event in context. A summary of the versions of the forecasting system used operationally and for ERA-40 is given in the next section. Section 3 provides an introductory discussion of forecasting-system performance in cases of similar vortex breakup and sudden warming in the Northern Hemisphere. Section 4 presents operational analyses and forecasts of the September 2002 event in the Southern Hemisphere. Computational stability is discussed in section 5, and reruns of the data assimilation and forecasts are discussed in section 6. One of the reruns is examined in further detail in section 7 in terms of the accuracy of the analyses and forecasts and the impact of the satellite and radiosonde components of the observing system. Section 8 reports on a search through the ERA-40 archives to check for any indications of similar behavior in earlier years. Secondary vortex development by barotropic instability is discussed in section 9, and concluding remarks are made in section 10.
2. The ECMWF forecasting system
The operational ECMWF system is based on a global atmospheric model with comprehensive parameterizations of physical processes, tightly coupled with an ocean wave model. In its highest-resolution application, the atmospheric model is run with a spherical-harmonic horizontal representation that since November 2000 has been truncated triangularly at wavenumber 511. The corresponding computational grid has an almost uniform spacing of just under 40 km. Variables are represented at 60 levels in the vertical, ranging from a height of 10 m above the surface to a pressure of 0.1 hPa (a height of about 65 km). Vertical resolution is approximately uniform in height with a spacing of about 1.5 km in the middle stratosphere between 70 and 3 hPa. This model is used for data assimilation and to produce deterministic high-resolution forecasts up to 10 days ahead. Ensembles of perturbed lower-resolution medium-range forecasts are also run operationally to provide probabilistic tropospheric predictions, using a model version with T255 spectral resolution and a 40-level vertical resolution. This version extends only to 10 hPa, and although the (unperturbed) control forecasts of 10-hPa height for the austral stratosphere in September 2002 from the ensemble system were reasonably skillful, they were generally less accurate than the 60-level operational forecasts. All operational forecasts presented here are from the higher-resolution version of the model.
An incremental four-dimensional variational data assimilation (4DVAR) is used to produce initial conditions for the medium-range forecasts. Corrections to 12-h high-resolution background forecasts for 0000 and 1200 UTC (analysis increments) are derived using a T159 resolution version of the model to optimize the fits of forecasts to in situ and remotely sensed observations taken in the time windows 1500–0300 and 0300–1500 UTC. Of the types of data used, temperature and wind measurements from radiosondes and radiances from the satellite-borne Advanced Microwave Sounding Unit-A (AMSU-A) instruments have the most direct influence on the middle stratospheric analysis, with the radiance data predominant in the extratropical Southern Hemisphere, as will be seen later. Radiance data are assimilated directly; changes to the background model fields are sought that minimize the mismatch between simulated model-based radiances and measured values. Data from all of the stratospheric sounding channels of AMSU-A are used.
Results will be presented from two versions of the operational forecasting system. The first, referred to as cycle 25rl, was operational in September 2002. The second, cycle 25r4, superseded cycle 25r1 for operational use in January 2003. The change included improvements to the physical parameterizations of the model, improvements to the formulation of the 4DVAR analysis, and the assimilation of several new types of satellite data. During September 2002, AMSU-A data from instruments on two National Oceanic and Atmospheric Administration (NOAA) satellites (NOAA-15 and -16) were assimilated operationally, while data from the newly launched NOAA-17 satellite were under assessment. NOAA-17 data have been used in operations in addition to data from NOAA-15 and -16 since late October 2002. They have been used also in two of the reruns for August and September 2002 discussed in this paper.
ERA-40 is a project in which observations made in the period from September 1957 to August 2002 have been reanalyzed using a 6-hourly three-dimensional variational version (3DVAR) of the ECMWF data assimilation system. The assimilating model used a coarser T159 spectral truncation but the full operational 60-level vertical resolution. The data assimilation was based on an earlier cycle (23r4) of the forecasting system operational in the second half of 2001, modified to include a few newer features subsequently used operationally in cycles 25r1 and 25r4. Production of the ERA-40 analyses was completed during April 2003. Ten-day forecasts from each of the 0000 and 1200 UTC ERA-40 analyses have subsequently been completed for each year from 1958 to 2001, employing the model version used for the ERA-40 data assimilation.
A substantial documentation of the operational and ERA-40 data assimilation systems, including extensive references, may be viewed on the ECMWF Web site (http://www.ecmwf.int). Reference will be made below to specific papers of particular relevance to the present study.
3. Analyses and forecasts of Northern Hemispheric vortex breakup
Fifty years prior to the remarkable September 2002 event in the Southern Hemisphere, Scherhag (1952) reported an equally remarkable and never-before-observed warming of more than 40°C over 2 days measured at heights above 30 km by radiosondes launched from Berlin late in February that year. Originally dubbed “The Berlin Phenomenon,” it subsequently became clear that such “sudden warmings” are by no means rare in the Northern Hemisphere winter stratosphere and can be associated with substantial changes in the large-scale circulation pattern. On occasion, these changes comprise simply distortion, growth or decay, or a spatial shift of the principal features of the wintertime stratospheric circulation, namely, the cold polar vortex and the Aleutian anticyclone. Once every few years, however, the vortex elongates and splits substantially or completely into two, often accompanied by development of a second anticyclone.
Three examples of such vortex breakup are presented here. These cases have not been specially chosen because they exhibit an unusually high level of forecast skill. Rather, two are well-known cases from the past that can now be reexamined using the ERA-40 analyses and forecasts, and the third is simply the latest occurrence, in February 2003, as captured by the operational ECMWF system. In each case, 10-hPa height analyses are presented for a date on which the vortex was split and for 5 and 9 days earlier, together with forecasts of the vortex split made 5, 7, and 9 days ahead. The 9-day range is the farthest ahead at which some indication of splitting, albeit weak, is seen in all three cases.
The first example from ERA-40, shown in Fig. 1, is from January 1958. Scherhag (1960) referred to this case as a recurrence of the Berlin Phenomenon after the passing of 6 yr. The 10-hPa analysis of geopotential height for 20 January shows a substantial cyclonic vortex centered well off the Pole between northern Norway and Greenland, accompanied by a quite strong Aleutian anticyclone. Four days later, the planetary wave pattern has rotated eastward, the vortex has assumed a more elongated and bowed shape, and high pressure has built over southern and eastern Europe. By 29 January, the vortex has split completely into two separate vortices, one centered over Canada and one over Russia. Anticyclones are located over the Bering Strait and southwest of Ireland. The evolution depicted in these 10-hPa ERA-40 analyses matches well that seen in 25-hPa analyses reported at the time by Teweles and Finger (1958) and by Scherhag (1960).
The 5-day forecast for 29 January is clearly successful in capturing the vortex split, the principal error being a slight eastward shift of the overall pattern, a shift evident also in the 7- and 9-day forecasts for this date. Apart from the shift, each of these forecasts represents quite well the twin anticyclones and main cyclonic center over Russia. The treatment of the weaker vortex over Canada is, however, increasingly poor with increasing forecast range. Nevertheless, all forecasts out to day 6 show complete splitting of the vortex, in the sense that no closed contour line encompasses both low centers in height maps such as those shown. The radiosonde measurements (which include soundings from fixed ocean weather ships) and conventional surface observations from 1958 evidently permit remarkably good forecasts of this major stratospheric warming, when processed by a skillful modern data assimilation system.
The second example from ERA-40, presented in Fig. 2, is the event that occurred in February 1979. Although there are differences in the size and orientation of the polar vortex as analyzed for 12 and 16 February 1979 compared with 20 and 24 January 1958, the subsequent vortex breakup results in a pattern on 21 February 1979 that is similar to that of 29 January 1958. The situation on 21 February can be forecasted accurately at the 5-day range, but there is a quite rapid decline in skill beyond 6 days ahead.
This 1979 event has been much studied (see Jung et al. 2001, and references therein). ECMWF first produced analyses for the period as part of its contribution to the First Global Atmospheric Research Programme (GARP) Global Experiment (FGGE). Forecasts run from the FGGE analyses by Bengtsson et al. (1982) and Simmons and Strüfing (1983) provided highly accurate depictions not only of the February vortex split but also of the major circulation changes that occurred in preceding weeks. This was the first time such forecast skill had been demonstrated for stratospheric warming events, although some degree of success had been reported earlier for pioneering experiments by Miyakoda et al. (1970) for the case of March 1965.
Since then, ECMWF’s operational forecasts have proved generally successful at predicting major warming events in the Northern Hemisphere, for up to a week or so ahead. Figure 3 illustrates operational performance for the most recent case, in February 2003, which resembles that of 1979 in the sequence of 10-hPa analysis maps shown. In this instance both 5- and 7-day forecasts accurately depict the two cyclonic vortices centered over Canada and Russia and the anticyclone over the Bering Strait. The forecasts are less successful in positioning the weak pressure minimum south of the Aleutian high on 17 February, a more significant feature when viewed in wind maps. Further discussion of this feature is given in section 9.
4. Operational forecasts of the September 2002 event in the Southern Hemisphere
The stratospheric vortex in the Southern Hemisphere split completely on 25 September 2002 at 10 hPa and above, as depicted in maps of the operational ECMWF analyses of geopotential height.
The top and middle plots of Fig. 4 show the 10-hPa height analyses for 15, 20, 25, and 30 September 2002. On the 15th and 20th, the cold vortex was already displaced off the South Pole toward the South Atlantic. An anticyclone was centered south of Western Australia.
Major changes occurred over the next 5 days. The vortex elongated, moved farther away from the Pole, and split into two. The anticyclone south of Australia intensified considerably, and a smaller anticyclone developed over the South Atlantic. In the course of the following 5 days, one of the two resulting cyclonic vortices, that over the South Pacific, weakened considerably and moved slightly westward. The other cyclone and dominant anticyclone moved eastward and poleward, the anticyclone continuing to build while the cyclone decayed.
This dramatic evolution of the southern stratospheric circulation was captured extremely well by the 10-day operational forecast made on 20 September. The bottom two plots of Fig. 4 show the 5- and 10-day 10-hPa height forecasts valid on 25 and 30 September, respectively. The 5-day forecast is in almost perfect agreement with the analysis. Moreover, the essence of the change from day 5 to day 10 is described, although errors of positioning and intensity not surprisingly become evident by the end of the forecast range.
The 7- and 10-day operational 10-hPa height forecasts valid on 25 September are shown on the left-hand side of Fig. 5. The 7-day forecast for 25 September is largely successful; at the time illustrated, vortex splitting is not quite complete, but this is due to a slight delay in timing rather than a failure to fully capture the split. The vortex is unusually elongated, but not split, in the 10-day forecast for this date. Corresponding 8- and 9-day forecasts show the formation of two cyclonic centers, but no complete split of the vortex. Complete vortex splitting occurs at 10 hPa in all forecasts made from 18 September onward, and at 7 hPa and above in all forecasts from 17 September onward.
The right-hand side of Fig. 5 shows the 5- and 7-day forecasts valid on 30 September. An improvement can be seen in the location and orientation of the main cyclone and anticyclone from day 10 (Fig. 4) to day 7. The remnant of the second cyclone is too weak and too far to the west over the Pacific at day 7, rather than too far to the east as it is at day 10. This feature is depicted well at the 5-day range.
5. A weak computational instability
The extreme flow conditions in the austral stratosphere in the final week of September 2002 exposed a small-scale computational instability of the forecast model that had not been seen previously in either test or operational use. The instability was self-limiting in that the operational data assimilation and forecast runs did not fail to complete. Nevertheless, considerable localized noise occurred in the analysis for 26 September, and noise was present to a lesser extent on the following day. This can be clearly seen in the 10-hPa height analyses for 26 and 27 September shown in the upper plots of Fig. 6. Noise develops in the region of strong easterly flow in each of the two vortices, first appearing close to where the flow lies above mountainous coastal regions of Antarctica. In Fig. 6, the noise has reached large amplitude in part of the Pacific vortex in the operational analysis for the 26th (labeled 25r1) and is present to a lesser extent in the other vortex over and near Queen Maud Land, Antarctica, on the 27th. The noise is highly predictable in that it occurs also in forecasts out to as far as 8 days ahead valid on these dates, indicating that it is linked to the well-predicted large-scale flow conditions.
Linear analysis of the stability of the model’s semi-Lagrangian advection scheme (Hortal 2002) shows that the scheme can indeed be unstable if the change in velocity along a single time step trajectory is sufficiently large. In the present case the cause of the problem has been found to be the combination of a relatively long trajectory associated with strong vortex flow and a substantial along-trajectory change in the vertical component of velocity associated with gravity wave motion. The gravity waves are generated by flow over the Antarctic coastal mountains, propagate upward, and amplify in the regions of easterly vortex flow. The amplification results in vertical velocities that change sufficiently along trajectories to trigger the computational stability. Using a stable, first-order scheme for the vertical part of the trajectory calculation removes the noise but degrades large-scale forecast accuracy when applied everywhere.
6. High-resolution reruns
Given the rarity of occurrence of the instability discussed above, one possible solution is to monitor the change in vertical velocity along the trajectory in the stratosphere, and to apply the first-order scheme only where a critical value is exceeded. This solution has been tested in a rerun at operational resolution of the data assimilation and forecasts from 1 September to 15 October 2002. In addition to modification of the advection scheme, cycle 25r4 of the forecasting system was used for the rerun as it had been tested and shown to be ready for operational implementation. The lower-left plot of Fig. 6 shows an almost completely noise-free 10-hPa height analysis for 26 September from this assimilation.
A second rerun of the forecasting system has also been carried out using cycle 25r4 at operational resolution. It covers the period from 1 August to 30 September 2002 and assimilated sounder data from the NOAA-17 satellite in addition to the data from NOAA-15 and -16 used operationally and for the first rerun. This second rerun was originally set up to test the three-satellite configuration with the new cycle of the forecasting system and to provide a control experiment for a set of observation denial experiments, rather than specifically for study of events in the southern stratosphere. In this case the time step was simply halved in the assimilation cycles beyond 1200 UTC 25 September to limit the computational instability. The lower-right plot of Fig. 6 shows the analysis for 26 September from this second rerun. Noise is much reduced compared to operations, but not to quite the same extent as found from changing the advection scheme. A similar result is found for the rerun analyses for 27 September (not shown).
Apart from the noise, differences between the operational analyses and the two rerun analyses are generally small, although both reruns can be seen to give slightly higher 10-hPa heights in the vicinity of the Antarctic Peninsula in the analyses for 26 September shown in Fig. 6. As would be expected, larger differences are evident at later ranges of the forecasts carried out from these analyses. An example is shown in Fig. 7, which presents 5- and 10-day 10-hPa height forecasts from the analysis for 20 September 2002 using cycle 25r4 and data from the three NOAA satellites. Compared to the operational forecasts presented in Fig. 4, there is little difference at the 5-day range apart from a very slightly earlier split of the vortex in the case of cycle 25r4. The 10-day forecast from cycle 25r4 is, however, clearly superior in this case. More generally, objective verification such as presented in the following section indicates a small but statistically significant benefit from the use of cycle 25r4 in the first half of the forecast range and a very modest additional benefit from the use of the additional data from the NOAA-17 satellite. Except where indicated, the results from cycle 25r4 presented in the remainder of this paper are based on the longer assimilation carried out using data from the three NOAA satellites, which is the configuration of the ECMWF forecasting system that became operational in January 2003.
Figures 8 and 9 present maps of the distributions of potential vorticity (PV) and water vapor on the 850-K isentropic surface. They complement the maps of 10-hPa height already shown. Figure 8 shows fields derived from the cycle-25r4 rerun analyses for 20 and 25 September 2002 and from the 5-day rerun forecast from 20 September. The analyzed vortex at this level is characterized not only by large negative values of PV but also by relatively high values of water vapor. The latter arise from the modeled descent of air from the upper stratosphere that has been moistened by a parameterization of methane oxidation in the background forecasts of the data assimilation. In the ECMWF assimilation system, background stratospheric values of PV are changed (predominantly on medium to large scales) by analysis increments to the wind and temperature fields, whereas background stratospheric humidity is unchanged by the analysis (apart from the rare removal of any supersaturation that results from changed temperatures). Quantitative values of humidity are open to question in the Tropics, where the slow upward transfer of dry air above the cold tropopause is not captured as well by the data assimilation as in the free-running simulations reported by Simmons et al. (1999). The contrast between a relatively moist polar vortex and dry Tropics is nevertheless at least qualitatively realistic, given that temperatures are not cold enough in the vortex to have caused condensation at this level or above.
The principal features of the PV and humidity fields are strikingly similar in pattern. Both fields change from day to day primarily due to advection by the distribution of winds on the isentropic surface. Small-scale structure is also introduced directly into PV (but not humidity) by the model’s parameterized orographic drag. This in particular modifies PV within the vortex (or vortices) over Antarctica.
Figure 8 shows clearly the well-predicted split of the vortex already seen in the height field maps of Fig. 4. It also shows multiple extended streams of large PV/moist air (darker-shaded bands) extruded from the vortex or vortices and streams of small PV/dry air (lighter-shaded bands) drawn in from low latitudes. The maps for 20 September depict in particular a band of air spiraling from low latitudes over the Pacific, across South America and the South Atlantic, toward Antarctica. Such a feature is seen repeatedly in the analyses for August and September. The minimum (in specific humidity and the magnitude of PV) over Queen Mary Land (around 90°E) on 20 September can be traced over the next 5 days to the location of the center of the intensified anticyclone south of Eastern Australia shown in Fig. 4. Indeed, throughout August and September, the growth and movement of anticyclonic features surrounding the vortex in the 10-hPa height analyses can be linked with air advected from low latitudes, as identified by distributions of PV and specific humidity on the 850-K surface.
Figure 9 shows corresponding analyses and 10-day forecasts for 30 September 2002. There is particularly strong extrusion from both vortices after the split on 25 September, especially from that over the South Pacific, consistent with the weakening of the vortices seen in height fields. Air of recent low-latitude origin almost surrounds each vortex. The 10-day forecasts differ from the analyses in detail but tell essentially the same story.
The structures seen in Figs. 8 and 9 are dynamically plausible; they indicate basic good behavior of the model’s semi-Lagrangian advection scheme at high horizontal and vertical resolution, and they link with the 10-hPa height fields that are generally well analyzed. It would nevertheless be of interest to verify them against independent satellite data, for example, the limb-sounded data from the Environmental Satellite (ENVISAT). In this context, the ozone fields produced routinely by the ECMWF system are further candidates for study. The ozone analyses exhibit the principal features on the 850-K surface shown for PV and specific humidity in Figs. 8 and 9, but small-scale structure is less evident, due presumably to a stronger effect of source/sink terms in the assimilating model. Maps of total ozone depict the vortex split, but a deficient parameterization of heterogeneous chemistry results in ozone hole depths that are underestimated by some 50 Dobson units compared with retrievals from Global Ozone Monitoring Experiment (GOME) and Total Ozone Mapping Spectrometer (TOMS) data. More successful ozone analyses and forecasts from a chemical transport model forced by ECMWF winds are reported by Eskes et al. (2005).
It is beyond the scope of this paper to investigate either the mechanism of the vortex split itself or the origin of the unusual conditions that led to its occurrence in 2002. The analyses produced by ECMWF and other numerical weather prediction centers are important resources for such investigations; the study by Charlton et al. (2005) provides an example of their use. Future studies that are to be based on ECMWF data should use the improved rerun analyses discussed here instead of the corresponding operational analyses. Access details are specified in the appendix.
7. Analysis and forecast accuracy
In this section, we further examine the accuracy of the cycle-25r4 results for August and September 2002 in the austral middle stratosphere.
Figure 10 presents time series of radiosonde observations of 20-hPa temperature from three stations located near the Antarctic coast, together with corresponding time series of analyzed and forecast values at the three locations. The forecast start times are 1200 UTC, and the ranges are 4.5 or 5 days (referred to as D + 5) and 9.5 or 10 days (referred to as D + 10), depending on whether the verifying observations are for 0000 or 1200 UTC.1 The level of maximum analyzed warming over the 5 days beginning at 1200 UTC on 20 September is 20 hPa, which is a standard reporting level reached by a reasonable number of radiosonde ascents. The locations of the three stations, Casey, Syowa, and Halley, are indicated by the labeled observations plotted on the uppermost map in Fig. 11.
Figure 10 shows large oscillations in temperature at Casey (66°S, 111°E) in August and early September, associated with distortion and movement of the vortex, whose cold core lies over the station at some times but not at others. For most of the rest of September, the vortex no longer overlies the station, as can be seen for the sample of analyses presented in Figs. 4 and 8. Temperatures in this later period are reported to be mostly a little above −40°C. The observations at Syowa (69°S, 40°E) reveal cold temperatures between about −90° and −80°C in the first half of August. Temperatures at this station subsequently fluctuate as the soundings become influenced by vortex changes in the period up until the last 10 days of September. Cold air (with temperatures mostly below −80°C and reaching as low as −97°C) lies above Halley (76°S, 27°W) until late September, when the observations (at 1200 UTC) show a rapid temperature rise of almost 60°C over the 6 days from 20 to 26 September as the vortex splits.
The analyzed values shown in Fig. 10 fit the radiosonde observations very closely, as do the forecasts at the D + 5 range, with just a few exceptions. Indeed, the major oscillations are captured quite well even at the D + 10 range. The largest discrepancy is at Halley, where the 10-day forecasts from 15 and 16 September fail to indicate the strong warming associated with the vortex split. The earlier forecasts for Halley show no sign of a systematic model drift in vortex core temperature over the 10-day forecast range. Figure 10 provides an indication that high levels of forecast skill are not unique to the late-September period.
Snapshots indicating a good fit to observations are presented in Figs. 11 and 12. Figure 11 shows contour maps of the 20-hPa temperature analysis for 1200 UTC 25 September and of the corresponding 5-day forecast for this time, with the available radiosonde measurements superposed. The change in temperature from 1200 UTC 20 September to 1200 UTC 25 September is also shown for the analysis and radiosonde measurements. A remarkable agreement is again seen between the analysis and the 5-day forecast, and both analysis and forecast match well the measurements from the limited number of radiosonde ascents. The maximum analyzed (and forecast) temperature rise exceeds 61°C over the 5 days; the nearby radiosonde measurements from Halley indicate a rise of 54°C.
A much denser observational coverage is provided by the soundings from polar-orbiting satellites. The plot on the left of Fig. 12 shows radiance measurements from channel 11 of the AMSU-A instruments on NOAA-16 and -17, accumulated over the assimilation period 0300–1500 UTC 25 September. Channel 11 ceased working on NOAA-15 earlier in 2002, but data from adjacent channels were available and assimilated for the period in question, providing coverage over the two blank triangular regions in Fig. 12 where no channel-11 measurement is available. The radiance measured by channel 11 is most sensitive to the atmospheric temperature at about 20 hPa, although it is influenced by values over a deep stratospheric layer, temperatures at 50 and 10 hPa having half the influence of temperature at 20 hPa. The distribution of channel-11 brightness temperatures is nevertheless in quite close agreement with the 20-hPa temperature analysis shown in Fig. 11.
The plot on the right of Fig. 12 shows corresponding differences between the measured radiances and radiances simulated using the 4DVAR background forecast from 0000 UTC 25 September. These differences generate a 1200 UTC analysis increment due to the channel-11 data. The differences are small, less than 1°C in brightness temperature at most points. Red-colored spots show where the measured radiances are most in excess of the simulated radiances. For the 12-h period illustrated, the channel-11 data tend to warm the analysis (or equivalently correct a slightly cold background field) west of the Antarctic Peninsula and over and to the east of Queen Maud Land. The data cool the analysis over the Atlantic sector of the Southern Ocean.
A selection of analysis and forecast verification statistics is presented in Fig. 13. Root-mean-square errors of 20-hPa temperature and vector wind are shown, evaluated over the extratropical Southern Hemisphere and averaged for initial analyses and forecasts run from 1200 UTC daily for the period from 1 August to 30 September 2002.
The top plots of Fig. 13 show verification of cycle-25r4 forecasts against subsequent cycle-25r4 analyses for 0000 and 1200 UTC and verification of the same forecasts and of the initial analyses against radiosonde measurements for these times. A diurnal cycle in the verification against radiosondes is due to a difference in the geographical distribution of the verifying observations at 0000 and 1200 UTC. If forecast error is not uniformly distributed over the extratropics, it is sampled differently by the radiosondes at the two verification times. During August and September 2002, over the region south of 20°S, the number of 0000 UTC radiosondes reaching 20 hPa varied between 19 and 35 each day, and the corresponding number of 1200 UTC radiosondes varied between 13 and 26.
Verification against analyses indicates lower forecast errors than verification against radiosondes out to about 5 days ahead. The verification against radiosondes exhibits slow error growth and little diurnal variation over this part of the forecast range.
Corresponding mean errors against radiosondes (not shown) are of the order of 0.5°C for temperature and 1 m s−1 for wind. These results suggest that initial analysis errors and consequential short-range forecast errors are much smaller than the errors of the verifying radiosonde observations. In this context, radiosonde observation errors comprise not only direct measurement errors, but also errors of timing and location (the whole balloon-based sounding is assumed valid at the single reported time and location) and representativity (if the measurement is influenced by small-scale motion unresolved in analyses and forecasts). The implied random analysis errors are of the order of a few tenths of a degree in temperature and 1 m s−1 in wind. The corresponding values for 5-day forecast errors are about 2°C and 6 m s−1. The small implied analysis errors for temperature are consistent with the accuracy of fit of simulated and measured AMSU-A brightness temperatures. The implied radiosonde observation errors of 1°C or a little more for temperature and about 4 m s−1 for wind at 20 hPa are reasonably consistent with the corresponding values of 1.5°C and 3.3 m s−1 specified in the data assimilation system. The specified observation error for the AMSU-A channel-11 brightness temperature is 0.35°C.
Quite similar analysis, forecast, and radiosonde observation errors are inferred from corresponding verifications for the 30-, 50-, and 70-hPa levels.
The middle plots of Fig. 13 show verifications against analyses comparing the cycle-25r4 forecasts with the operational (cycle 25r1) forecasts. The improvement of cycle 25r4 over cycle 25r1 out to 5 days ahead (rather larger for wind than temperature) is statistically significant at the 0.1% confidence level for a t test applied to the wind score differences, assuming them to be temporally uncorrelated. The superiority of the newer cycle is clearly confirmed by verification against radiosondes in the case of the wind field (not shown). This supports the recommended use of the high-resolution rerun analyses rather than the operational analyses for further studies. The poorer temperature scores for cycle 25r4 beyond day 7 are of low statistical significance.
The bottom plots of Fig. 13 show verifications against 1200 UTC radiosondes of the standard 25r4 forecasts and of forecasts from assimilations in which either all satellite data or all radiosonde data2 were withheld throughout the August and September period. Withdrawing radiosonde data from the assimilation not surprisingly results in analyses that do not fit the radiosonde data as well, as shown by the day-0 differences in the plots. There is, however, only a very slight degradation of subsequent forecast quality from this data denial. Five- and ten-day 10-hPa height forecasts run from the no-radiosonde analysis for 1200 UTC 20 September are shown in Fig. 14 and can be seen to be similar to the cycle-25r4 control forecasts shown inFig. 7.
In contrast, removing the satellite data causes a very substantial degradation of the fit of analyses and forecasts to the radiosonde data. The analysis for 25 September (not shown) represents quite well the strong anticyclone south of Eastern Australia but produces a vortex that is unusually elongated rather than split, similar to that displayed in Fig. 5 for the 10-day operational forecast valid for this day. It should be stressed that the no-satellite assimilation used the same background error statistics as the standard cycle-25r4 assimilation. Better use of the limited available radiosonde data could have been made by specifying larger background errors (as was done to a degree for the ERA-40 system). The present results nevertheless demonstrate the overwhelming role played by satellite data in determining the high quality of analyses and forecasts for the Southern Hemisphere stratosphere produced by the current ECMWF forecasting system.
8. A search through the ERA-40 archives
When the Southern Hemisphere vortex broke up in September 2002, it was thought to be an event that had never before been observed. For confirmation, 10-hPa height analyses from ERA-40 have been examined for each day in September and October from 1957 to 2001.
No instance of pronounced vortex splitting similar to that of 2002 has been found. It is not uncommon to see a marked weakening of the vortex toward the end of October, and the weakening vortex may be displaced from the Pole as a prelude to the establishment of summertime easterlies. This commonly occurs in conjunction with formation of a relatively strong anticyclone in the Australian sector, with the vortex displaced into the Atlantic/American sector, as indeed occurred prior to the vortex split in September 2002. This displacement tends to keep cold air over the Halley radiosonde station. This suggested examining all 20-hPa radiosonde temperature observations from Halley processed by ERA-40 for the months of September and October from 1957 onward to see how unusual the warm temperatures reported at Halley in late September 2002 were. No observation was found that was warmer than the value of −16°C measured in the ascent from Halley on 28 September 2002, the maximum value shown in the time series presented in Fig. 10.
The 20-hPa temperature reports from all other Antarctic stations located poleward of 70°S have also been examined for the months of September and October from 1957 to 2002. The warmest temperature in 2002 was −10°C measured at Neumayer (71°S, 8°W), also on 28 September. This is higher than any earlier 20-hPa temperature in the ERA-40 database for this set of stations and the 2 months in question. The previous high values of −11°C from Vostok (78°S, 107°E) on 14 October 1979 and −12°C from the South Pole on 28 October 2000 were each associated with displacement of the (single) analyzed vortex away from the Pole into the Atlantic/American sector. A more comprehensive discussion of the radiosonde record that exists from the late 1950s onward, supplemented by results from ECMWF’s earlier ERA-15 reanalysis, is given by Roscoe et al. (2005) in this issue. Consistent with our results, Roscoe et al. find no indication in the observational record of a vortex split prior to final warming, other than that in September 2002.
Comment must be made on biases in southern polar stratospheric temperatures in the ERA-40 analyses. Mean temperatures for late winter and early spring for the pre-1973 period, when no satellite data and only particularly sparse radiosonde data were assimilated in the stratosphere, are typically colder than those for later years by 10°C or more south of 60°S between 20 and 50 hPa. A similar bias was reported by Simmons et al. (1999) for an earlier version of the model run in climate simulation mode. The early analyses are thus liable to overestimate the intensity of the austral polar vortex and may not provide a reliable representation of perturbations to it. Temperature biases are smaller when satellite data are assimilated, but the later ERA-40 analyses exhibit an oscillatory temperature structure in the vertical with an amplitude of a few degrees throughout the Antarctic stratosphere. This is seen neither in other climatologies (Randel et al. 2004) nor in corresponding operational analyses. Biases against Southern Hemisphere radiosonde observations are accordingly much larger than for the operational analyses. A problem in the 3DVAR assimilation of radiance data in ERA-40 that is not experienced to the same degree in the operational 4DVAR system is the likely cause. This has little impact on analyzed synoptic characteristics, which are very similar in ERA-40 and operations for the 1999–2001 period when both assimilations used similar vertical resolution in the stratosphere. It nevertheless complicates the interpretation of time series of analyzed Antarctic temperatures.
9. Secondary vortex development
Searching through the ERA-40 archives did reveal a number of cases in which the analyzed vortex elongates considerably and distorts but does not break up into two vortices of similar intensity. One such case was found for late October 1972, at a time when the 20-hPa temperature measurement from Halley reached −19°C, the highest September or October temperature for this station at this level prior to September 2002, according to the ERA-40 database.
Figure 15 presents a more recent example that serves also to illustrate additional diversity in vortex dynamics. The figure shows maps of 10-hPa height and 850-K PV and specific humidity for 21, 23, and 25 October 1994. On 21 October, the vortex is highly elongated and bowed, and flanked by two anticyclones. The height map for 2 days later shows a second low center in the main vortex, but this does not develop further. There is also a weak low pressure center cutoff from the trailing portion of the main vortex, located west of South America. This cutoff low subsequently moves westward around the Pacific anticyclone and intensifies. The PV and humidity maps for 25 October each depict cyclonic wrapping up near the end of a band of material extruded from the main vortex, in a manner similar to that seen in idealized models of vortex dynamics (e.g., Polvani and Plumb 1992). The picture is somewhat sharper for humidity, which is a prognostic model gridpoint variable, than for PV, which is derived from the model’s spectrally represented prognostic dynamical variables. The left-hand plot in Fig. 16 presents a local map showing the distributions of Montgomery potential and wind for the developing perturbation on the 850-K isentropic surface. The tilt of the system is counter to the shear of the ambient flow, indicating that barotropic instability of the local easterly flow around the Pacific anticyclone plays a role in the intensification.
Hartman et al. (1996) indicated the possibility of such perturbation growth by applying a singular vector analysis to a case of extrusion of high-PV air around the Aleutian anticyclone in the wintertime Northern Hemisphere. An example of this is provided by the weak low near the date line in the analysis for 17 February 2003 shown in Fig. 3. The right-hand plot in Fig. 16 shows it to have a structure in its developing phase that is very similar to that of the October 1994 case (taking into account the basic hemispheric difference).
The same process also appears to be responsible for breakup of the extrusion from the weaker of the two Southern Hemisphere vortices at the beginning of October 2002, following the vortex split one week earlier. Figure 17 displays contour maps of 10-hPa height superimposed on the distributions of specific humidity on the 850-K surface for 1200 UTC 2 and 4 October 2002. These analyses are from the cycle-25r4 assimilation with modified advection scheme. The figure shows how the original Pacific vortex and band of air extruded downstream of it (depicted in Fig. 9 for 30 September) breaks up into three synoptic-scale vortices, as features move westward around the strong anticyclone. The map for 2 October also shows a streamer of relatively moist air that is drawn poleward from just south of Australia and wrapped around the primary vortex. This air was extruded from the weaker Pacific vortex several days earlier.
Comparisons with observations are needed to validate analyses of secondary vortex developments such as those discussed above. These developments tend to take place over regions with particularly sparse coverage by in situ measurements, especially at the heights in question, and are likely to be less well treated by the assimilating forecast model than are forced planetary-scale features. The analyzed synoptic-scale systems illustrated for early October 2002 have too shallow a vertical structure to be resolved well by AMSU-A measurements, but they move over a region of relatively good in situ data coverage. The data assimilation for this period fits well the variations in 10-hPa temperature and wind measured by Australian and neighboring radiosondes. In contrast, the secondary vortex near the date line on 17 February 2003 is a relatively deep feature that increases in intensity above 10 hPa and can therefore be “seen” by satellite soundings. Its thermal pattern closely matches the pattern in the assimilated radiances from the high-sounding AMSU-A channels.
Features in analyses that are directly forced at small scales by the assimilating model must be viewed with caution. The PV map for 23 October 1994 presented in Fig. 15 shows U-shaped bands of low and high PV extending from the Antarctic Peninsula into the South Atlantic and then across southern South America and into the Pacific. These bands arise from advection of vorticity forced persistently in a dipole pattern over the Antarctic Peninsula on 22 and 23 October by the model’s parameterization of gravity wave drag. No such feature occurs in the corresponding humidity field. The bands spread to the central Pacific by 25 October, by which time a new dipole feature in PV (but not humidity) has formed downstream of the Antarctic Peninsula.
10. Concluding remarks
The splitting of the austral polar vortex at 10 hPa in September 2002 was an event the like of which had not been previously observed in the Southern Hemisphere, yet it was predicted a week or so in advance by ECMWF’s operational forecasting system.
Examples have been presented here of the successful prediction of such events in the Northern Hemisphere, a capability for which was first demonstrated more than two decades ago. The accuracy of stratospheric forecasts has been enhanced in recent years by improvements in observations (with the AMSU-A data available from 1998 onward), in data assimilation methods [e.g., through introduction of variational techniques (Andersson et al. 1998; Rabier et al. 2000) and direct radiance assimilation (McNally et al. 2000)] and in modeling [such as ECMWF’s introduction of 60-level resolution (Untch and Simmons 1999) and finite-element discretization in the vertical (Untch and Hortal 2004)]. Moreover, medium-range forecasts for the Southern Hemisphere troposphere have been brought to levels of accuracy similar to those reached for the Northern Hemisphere troposphere (Simmons and Hollingsworth 2002). Given that flow conditions in September 2002, however unusual, were conducive to the occurrence of a major breakup of the Southern Hemisphere vortex, it would have been surprising had the forecasting system failed to predict this breakup well in advance.
The accuracy of the forecasts discussed in this paper is indicative of the high quality of the analyses from which they are run. Analysis error is estimated to be substantially less than radiosonde observation error. A strong control on the large-scale stratospheric analysis is provided by satellite radiance data, and further control may arise from the data assimilation system’s upward propagation of information from the better-observed troposphere. Although radiosonde data do not play a strong direct role in determining the quality of the stratospheric analyses and forecasts for the period examined, they provide valuable information for validation of the forecasting system in the stratosphere. They are also important in determining the bias corrections that are applied to the satellite radiance data prior to assimilation. Their importance as part of the long-term stratospheric climate record has been illustrated by our use of the data to indicate the extreme nature of the warming of late September 2002. Moreover they have a larger beneficial impact on the quality of forecasts for the troposphere than shown here for the stratosphere.
The ERA-40 analyses provide a comprehensive description of the stratosphere from August 1957 onward and have been used here to help place the events of September 2002 in historical context. Although these analyses are applicable for a wide range of studies, they should be used with care where observational data are sparse and the assimilating model is prone to systematic error. This is particularly the case for the Northern Hemisphere upper stratosphere, and for the Southern Hemisphere more generally, prior to late 1978, when data first became available from the stratospheric-sounding channels of the Microwave Sounding unit (MSU), High Resolution Infrared Sounder (HIRS), and Stratospheric Sounding Unit (SSU) instruments on the operational NOAA polar-orbiting satellites. Analyses below 10 hPa in the Northern Hemisphere are much less sensitive to changes in the observing system, and those for January 1958 have been shown to be of sufficient quality to enable good forecasts of the major warming that occurred late in that month.
We are grateful to Ernst Klinker for providing information on the work of Scherhag; to Tim Palmer for reminding us of the work of Hartmann et al.; and to Horst Böttger, Tony Hollingsworth, and two reviewers for comments on the text. The ERA-40 project was partially funded by the European Union under Contract EVK2-CT-1999-00027 and was supported by Fujitsu Ltd through the provision of computational resources.
Access to the Rerun Analysis Data
Readers with direct access to ECMWF’s archives can retrieve analysis data at 6-hourly intervals from 0000 UTC 1 August to 1800 UTC 30 September 2002 from the cycle-25r4 rerun including assimilation of data from NOAA-17 by specifying CLASS = RD, EXPVER = “ec7s”, TYPE = FC, STEP = 0 in their Meteorological Archive Retrieval System (MARS) retrieval statement. To retrieve data from the rerun from 1 September to 15 October using the more stable advection scheme (without NOAA-17 data), specify EXPVER = “ecal”. Readers without direct access can order data from ECMWF Data Services via the ECMWF Web site (www.ecmwf.int). A specific dataset that can be supplied for research use at a nominal handling charge is available for the period 1 August–15 October. It is based on the first of the above reruns for the month of August and on the second for 1 September–15 October.
Corresponding author address: Dr. Adrian Simmons, European Centre for Medium-Range Weather Forecasts, Shinfield Park, Reading RG2 9AX, United Kingdom. Email: email@example.com
The times given in radiosonde reports vary from station to station. Reports timed at either 2300 UTC the previous day or 0000 UTC are regarded here as observations for 0000 UTC, and 1100 and 1200 UTC reports are likewise both regarded as observations for 1200 UTC.
The bias corrections that were applied to the satellite data used in the no-radiosonde assimilation were derived from earlier assimilations that had used radiosonde data, following Harris and Kelly (2001). Pilot balloon and wind profiler data were also withheld in the no-radiosonde assimilation.